OpenAI has recently unveiled its latest advancements in image detection technology. In a recent blog post, the company revealed its efforts to enhance the provenance methods for tracking and verifying AI-generated content.

One of the key highlights of OpenAI’s announcement is the introduction of a new image detection tool designed to identify images created using its DALL-E AI image generator. This tool employs advanced AI algorithms to determine the origin of a given image, even in cases where the image has been altered through cropping, compression, or changes in saturation.

In addition to the image detection tool, OpenAI has also implemented new watermarking methods to effectively label and track content generated by its AI systems. These tamper-resistant watermarks serve as invisible signals that can be embedded in various types of content, including audio files.

OpenAI’s image detection classifier demonstrates impressive accuracy in identifying images produced by DALL-E 3, boasting a success rate of around 98%. However, its performance in recognizing content generated by other AI models, such as Midjourney, is currently less robust, flagging only a small percentage of images from these sources.

OpenAI’s commitment to content provenance is further underscored by its collaboration with the Coalition of Content Provenance and Authority (C2PA). By incorporating content credentials into image metadata, OpenAI aims to provide transparency regarding the ownership and creation process of AI-generated content. The company’s active participation in C2PA, including its recent inclusion in the steering committee, highlights its dedication to industry standards and best practices.

While the image classifier and audio watermarking technology are still undergoing refinement, OpenAI is eager to gather feedback from users to assess their effectiveness. Interested parties, including researchers and nonprofit organizations, can access and test these tools through OpenAI’s research access platform.

This article was originally published on May 7, 2024, at 5:40 pm ET.